AI · November 16, 2026 · Low impact
AMD Loom adds Radeon 9700 support and reports a 42% average speedup over llama.cpp across 48 models
A developer added Radeon 9700 support to AMD Loom, an experimental inference stack that includes a runtime, kernel compiler and inference server. The project benchmarked 48 models with an average 42% speedup over llama.cpp. Earlier, on Strix Halo, the top 51 trending Q4_K_M models from Hugging Face averaged about 40% faster across decode and prefill, using the better of the ROCm or Vulkan backends. The compiler and runtime can also be integrated into other inference projects, and a llama.cpp backend is in progress.
It points to meaningful speed gains for local inference on AMD consumer GPUs, with the compiler reusable in other projects.